What Operators Reveal

The correction that vanished by the next session

Claude produced a tone classification that was exactly wrong — aggressive when the email was apologetic. Three exchanges to correct it: reframing, specifying, finally showing the model the right output. By the end, it worked. Three days later, a new session, a similar email, and the same wrong classification was back. What had vanished was the correction, not the model's behavior — the model did exactly what it was designed to do, which is forget everything once the context closes.

That's a capture failure, and it's common enough to be a background condition of intensive Claude use. Every session where Claude gets something wrong, and you fix it, produces a data point describing how your domain works and where the model's defaults diverge from your actual needs. Most of that signal evaporates. The thesis is stated plainly in the book: every correction is a data point. Undocumented, it's a one-time fix. Documented, it becomes a constraint that holds across sessions and compounds over months.

This book builds the correction-collection system that makes the difference: a flat JSONL store seeded by hand, organized by drift category, with entries promoted one at a time once a correction is solid enough to become a rule. Seven named Claude workflow patterns cover every stage, from recognizing drift mid-session to reading the full corpus as a capability assessment across model versions.

Turning fifteen months of corrections into a queryable map

• Recognizing drift before it compounds — catching a correction pattern early instead of re-discovering it session after session.
• The correction format — what makes a captured correction usable months later instead of a vague note nobody can act on.
• Building the collection system — a flat JSONL store, seeded by hand, organized by drift category.
• Querying what you have — turning an accumulating corpus into answers about where your domain diverges from Claude's defaults.
• Corrections becoming content — the promotion pipeline that turns a proven correction into a durable constraint.
• Reading the behavioral map and signal quality — a practitioner-level capability assessment built from your own accumulated evidence, across model versions.

A preview

The irony of leaving a successful correction undocumented is that you did the hard part. You figured out what was wrong, found language that fixed it, and then discarded all of it. Re-earning ground you already won is an expensive hobby.

Operators who've made the same correction more than once

This book is for practitioners running Claude intensively across real projects, who want their accumulated experience to compound instead of evaporating every time a session ends. It assumes no existing correction-logging system — the book builds one from a seeded corpus of ten entries upward. It does assume you've already noticed the pattern: re-teaching Claude the same lesson, more than once, because nothing from the last time you fixed it carried forward.

A longer excerpt is available to newsletter subscribers.

Buy Ebook on Amazon